Google’s Veo 3 has quickly become a test case for a harder problem in generative AI: what happens when high-quality video tools are used to package hate speech for social platforms built around rapid sharing?
According to a report from MediaMatters cited by Ars Technica, numerous TikTok accounts have posted AI-generated videos using racist and antisemitic tropes in recent weeks. The videos bear the “Veo” watermark and last up to eight seconds, tying them to Google’s video generator.
What is spreading on TikTok
The videos described in the report target several groups, with most of the content aimed at Black people. Ars Technica says the clips depict Black people as “the usual suspects” in crimes, absent parents, and monkeys associated with watermelon. The same wave of content also targets immigrants and Jewish people.
The short length matters because these clips are built for social media circulation. Eight seconds is enough time to deliver a stereotype, provoke a response, and invite comments. MediaMatters reported that original posts had many comments repeating the stereotypes shown in the videos.
Ars Technica described some examples as shocking and noted that shock itself may be part of the strategy. Anger and drama can drive engagement, and hateful AI content can be engineered to exploit that pattern with very little production effort.
Why Veo 3 changes the risk
Google released Veo 3 in May, and Ars Technica describes that release as a major jump in AI video quality. The same realism that makes harmless viral clips more convincing can also make abusive material more effective and easier to distribute.
AI-generated hate content is not new. The source article notes that people have used generative AI to create inflammatory and racist material for as long as the technology has existed. What changes with a more realistic video model is the polish, speed, and scale of the output.
Ars Technica also says it tested simple prompts with Veo 3 and found it easy to reproduce elements of the videos. That is a key enforcement problem: even if a model refuses direct requests, vague prompts or coded imagery can still lead to content that violates the spirit of the rules.
The source points to one example of that difficulty: racist tropes may appear indirectly, such as through the use of monkeys instead of humans in some videos. A guardrail can block obvious hate speech and still miss a visual reference that relies on context.
Rules exist, but enforcement is uneven
TikTok’s community guidelines prohibit this type of material. The source quotes TikTok’s rules as saying, “We do not allow any hate speech, hateful behavior, or promotion of hateful ideologies. This includes explicit or implicit content that attacks a protected group.”
On paper, that policy covers the racist caricatures described in the report. In practice, the videos still appeared and spread widely enough to become visible to users and researchers.
TikTok says it relies on both technology and human moderators to find content that breaks its rules. But the source also notes that the volume of uploads makes fast moderation difficult. That creates a familiar gap: a platform may ban content clearly, yet still fail to catch it before it gains attention.
A TikTok spokesperson told Ars Technica that more than half of the accounts cited in the MediaMatters report had already been banned for policy violations before the report was published. The spokesperson also said the remaining accounts have now been removed.
Google also has a Prohibited Use Policy that bans use of its services to promote hate speech, harassment, bullying, intimidation, and abuse. Ars Technica says the videos uncovered by MediaMatters appear to fall into one or more of those categories.
The moderation challenge is moving across platforms
TikTok is a natural place for short AI videos to spread because it is a major social video platform. But the source makes clear that the problem is not limited to TikTok.
Ars Technica notes that X, formerly Twitter, has gained a reputation for very limited moderation, contributing to a rise in hateful AI content. The article also says the issue could worsen because Google plans to integrate Veo 3 into YouTube Shorts.
That planned integration matters because it could place the same video-generation capability closer to another short-form video system. If similar content can be created and posted with fewer steps, platforms may face even more pressure to detect abuse quickly.
The central issue is not whether companies have written policies. TikTok and Google both do. The issue is whether those policies can stop harmful content before it is generated, uploaded, shared, and reinforced by comments.
What this reveals about AI guardrails
The Veo 3 videos show the limits of a safety model that depends on both AI guardrails and platform moderation. Google’s system is expected to prevent misuse at creation. TikTok’s system is expected to remove policy-breaking content after upload. The report suggests both layers can fail.
That failure does not require a total absence of rules. It can happen when prompts are vague, when imagery carries coded meaning, when moderators face too much content, or when engagement pushes offensive material before enforcement catches up.
For users, publishers, and platforms, the lesson is straightforward: realistic AI video raises the stakes for content moderation. A convincing clip can be made quickly, labeled with a watermark, and still circulate long enough to cause damage.
Ars Technica says it reached out to Google about Veo 3’s safety features but had not yet received a response. Until stronger prevention and faster enforcement are visible, the spread of racist AI videos remains a warning about the practical limits of today’s AI safety systems.